parallel

parallel runs a fixed, pre-defined list of agents simultaneously. All workers receive the same input. Results are collected into a map before the next step.

When to use it

• Independent analyses of the same input (financial, reputational, technical)
• Voting or consensus between multiple specialized agents
• Running different models or prompting strategies simultaneously
• Any task where you want multiple independent perspectives

AGENT.md

---
name: due-diligence
description: Runs financial and reputational analysis in parallel.
version: "1.0.0"
pattern: parallel
workers:
  - finance-analyst
  - reputation-analyst
  - legal-analyst
call:
  model:
    role: thinker
---

No body needed — the parallel pattern doesn’t make its own LLM call. It delegates to the listed workers.

Scaffold

tama add parallel my-agent

How it works

1. All workers start — every agent in workers: receives the same input and runs concurrently

2. Results collected — the runtime waits for all workers to complete

3. Output combined — all results are returned as a JSON map { "agent-name": "result", ... } with key="done"

Difference from scatter

	parallel	scatter
Workers	Different agents, heterogeneous	Same agent, homogeneous
Input per worker	Same input for all	Different item for each
Worker list	Static, defined at design time	Dynamic, decided by map agent at runtime
Reduce step	None (just collects)	reduce.md synthesizes results

Example: combining with fsm

parallel by itself just collects results. To do something with them, nest it in an fsm:

---
name: investment-analyzer
description: Runs parallel analyses then synthesizes a recommendation.
version: "1.0.0"
pattern: fsm
initial: analyze
states:
  analyze: synthesize    # parallel step
  synthesize:            # terminal
---

# agents/analyze/AGENT.md
---
name: analyze
description: Runs parallel due diligence analyses.
version: "1.0.0"
pattern: parallel
workers:
  - finance-analyst
  - reputation-analyst
---

# agents/synthesize/AGENT.md
---
name: synthesize
description: Synthesizes parallel analysis results into a recommendation.
version: "1.0.0"
pattern: oneshot
call:
  model:
    role: thinker
---

You receive a JSON map of parallel analyses.
Synthesize them into a single investment recommendation: BUY, HOLD, or SELL.
Justify your recommendation with specific points from each analysis.